Research On 3D Facial Expression Synthesis Based On Action Unit Feature Decomposition

With the rapid development of fields such as machine learning and humancomputer interaction,enabling smart devices to understand human emotions has become an important research topic.Among them,facial expression recognition for images is a basic and important research topic.At present,the expression recognition model based on the deep neural network has achieved high accuracy in the recognition task of global expression from face images.However,in order to accurately capture the emotional information by facial expressions,the model needs to pay more attention to the local facial expressions.Since it is difficult to obtain the local expression data with labels,a method of synthesizing data is usually used to generate a large amount of facial data with a specific expression.Face data usually exists in two forms,two-dimensional face images and threedimensional face models.Because two-dimensional face images lack the depth information of the face,3D facial expression data synthesis has gradually received more and more attention.At present,the commonly used 3D face generation models are based on the traditional 3D Morphable Model(3DMM)and divide the face shape information into identity information and expression information.The expression parameters in this model are a global representation of the facial expression but cannot control the local expression of the face.To ensure the control of the local expressions of the generated 3D face,we introduce the knowledge of the Action Unit(AU)that describes the relatively independent muscle movements of the face and decompose the global expression parameters to obtain the motion unit parameters as the local expression.This paper introduces a linear model and non-linear model to realize the generation of 3D facial expressions based on the parameter of AUs.Based on the prior knowledge of traditional 3DDM and AU,the linear model captures the muscle movement represented by a single AU from 3d face dataset and parameterizes it as the parameter of a single AU of the 3D face.In this way,the global expression parameters of the 3D face are decomposed into multiple AU parameters to realize the control of generating the local expressions of the 3D face.The biggest problem with the linear model is that the variety of 3d faces they generate depends on the size of the dataset.Therefore,we further propose a non-linear model.The non-linear model can generate more diverse 3D faces by a deep neural networks.Moreover,the process of decomposing global expression parameters into AU parameters is also using a network model to ensure that the decomposed features of multiple AU are relatively independent.After that,we proved through a large number of experiments that both models can control the local expression of the3 D face through the AU parameters.In addition,the experiment also proves that the decomposed parameters can be used as the representation of the action unit.